Surface and image-based registration methods with statistical modeling for biomedical applications

Over the past decade, digital data generation and collection has become increasingly important in biomedicine. Surgeons heavily rely on biomedical data for diagnoses, pre-operative planning, follow-up, etc. Learning from large collections of data, such as optical surface scans and images, can help us in automating diagnoses and reducing human subjectivity. The knowledge of shape variability in a large dataset can be a guide for product development for example. Letting a computer “understand” images based on past examples enables computer-assisted robotic surgeries. This thesis provides data-driven solutions for biomedical problems. On a first level, we present a framework to combine the shape information of different patients into one digital model. On a second level, these digital models serve as prior knowledge for computers to automatically “understand” new data. A crucial step on both, the modeling and the application level, is the anatomical alignment of data. This step, known as “registration”, involves the identification of corresponding points between different data which remains a challenging task to computers. The manuscript is divided into three parts. Part I provides an introduction to geometry and image processing, X-ray imaging and deep-learning. Part II presents the contributions of this thesis to (statistical) shape modeling of articulating bodies. Articulating statistical shape models (SSM) describe any individual up to a certain accuracy, while maintaining the possibility to be articulated into different poses. Hence, the acquisition of person-specific 3D models is no longer required. Two different articulating SSM’s have been constructed: a SSM of the human hand for splint design based on low quality 3D-scans, and a SSM of a horse limb for veterinary applications. The ability of SSM’s to describe many individuals also allows it to generate virtual data to train deep-learning models on. Part III of the thesis focuses on solving a specific registration problem in X-ray imaging. X-ray imaging or radiography is the most common imaging procedure for many orthopedic interventions thanks to its ability to visualize internal structures with a relatively low radiation dose and low acquisition cost. However, interpretation from 2D radiographs can be hampered by overlapping structures, magnification effects and the patient’s positioning. To avoid the difficulties associated with 2D projections, we developed two deep-learning methods, with and without statistical prior, to register a 3D model to a pair of radiographs. The registered model enables a 3D-interpretation, while keeping the benefits of RX over CT, in terms of costs and radiation dose.

Publication year:2023

Keywords:Doctoral thesis

Accessibility:Open